Optimizing the cloud? Don’t train models. Build oracles!

This repository contains the prototype of Cloud Oracles proposed in CIDR'24. See below for implementation and experiment details for precomputing and querying cloud oracles.

@inproceedings{cidr_2024/cloud_oracle,
  author       = {Tiemo Bang and Conor Power and Siavash Ameli and Natacha Crooks and Joseph M. Hellerstein},
  title        = {Optimizing the cloud? Don’t train models. Build oracles!},
  booktitle    = {14th Annual Conference on Innovative Data Systems Research, {CIDR} 2024 Chaminade, USA, January 14-17, 2024},
  publisher    = {{www.cidrdb.org}},
  year         = {2024},
  url          = {{https://www.cidrdb.org/cidr2024/papers/p47-bang.pdf}}
}

Precomputing the Cloud Oracle

The Hydroflow program in the precomputation directory demonstrates how to precompute Cloud Oracles for fault-tolerant object placement.

Specifically, this program implements the enumeration and filtering logic of §3.1 for: Placement of an object in 2 data centers 200km apart that minimizes the latency for writing to both data centers and reading from either.

Note that, this prototype uses synthetic data!

There are several variants. Monolith - Enumeration Only implements the logic of enumerating placement decisions and applying the distance constraint. Monolith additionally filters the enumerated decision by computing which decisions have dominated access latency and hence are never a good choice. Both these variants are monolithic in that they yield highly optimized but single-threaded executables. A parallel variant, which will execute the precomputation in data-parallel and pipelined stages, is pending -- see parallel branch.

Setup

Install the rust, e.g., via rustup (for Linux/Mac:curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh).

Run the experiment

cd precomputation
cargo run -r -- --output_dir=../results

This command compiles Hydroflow program in release mode, executes the experiment, and write the result in to results/precomputation/precomputation.csv.

The experiment computes a number of Cloud Oracles for different numbers of data centers hosting object stores (num_d) and data centers hosting client applications that access the objects (num_c). These parameters are hard-coded in main.rs according to the experiments in the CIDR publication.

Querying the Cloud Oracle

The python package in oracle_python implements several online optimization tasks of cloud oracles: scenario planning via simulation, and migration planning under workload drift.

The query logic on top of the precomputed look-up structure is implemented in queries as specified in §3.2, §3.3, §3.4 of the paper, respectively. The experiments in experiment are specified likewise. In the subdirectories, you will find for each use case the experiment definitions in use_case_X.py and the setup logic in experiment_X.py.

Note that, the experiments use synthetic data!

Setup

cd oracle_python
python -m pip install .

Run the experiments

python -m cloud_oracle_prototype --all --output-dir results

This executes all experiments of the CIDR publication. Alternatively, experiments can be run individually, see python -m cloud_oracle_prototype --help. The GPU- or MPS- based experiments are only executed if the according hardware is detected.

Plotting results

See results.ipynb for plotting. It assumes all result are stored the in results directory and according subdirectories of the experiments.

Setup

Assuming the basic infrastructure for Notebooks is present, install the dependencies for plotting:

python -m pip install -r requirements.txt